README.md

TaskTrove model data quality

This directory is the single experiment root for the five-model TaskTrove v3.42 reward comparison and its supporting data-quality work.

Layout

• TRACKER.md: experiment-level status and links
• reports/: unified reward reports and harness comparisons
• campaigns/: model-specific trackers, state, manifests, results, and launch tooling
• findings/: sampled rollout findings and remediation logs
• audits/: dataset-quality and verifier audit reports
• scripts/reports/: shared report generators
• scripts/tasktrove-repairs/: repair, validation, and publication tools
• artifacts/tasktrove-repairs/: repair inputs, staged datasets, and fixtures
• tooling/: pinned external tooling checkouts used by the campaigns

Each campaign is self-contained under campaigns/<campaign>/: its tracker is at the campaign root, executable control code is in scripts/, and durable local state is in artifacts/.